Retraining of open models also drops the cost of small fauna pretty dramatically, orthogonally to hardware changes, so further innovations there may drive a lot too. There's also the benevolent bot-farm approach (ala SETI@home) that might help parts of the fringe keep up.

There was a lot of algorithmic optimization research over the last few years which I see appearing in papers, meaning reduced computing needs. Combined with smaller, narrower domain models, I think the need for Big is deemphasized